-
Notifications
You must be signed in to change notification settings - Fork 5
Expand file tree
/
Copy pathfla_tuning.py
More file actions
294 lines (285 loc) · 7.73 KB
/
Copy pathfla_tuning.py
File metadata and controls
294 lines (285 loc) · 7.73 KB
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
import importlib
from typing import Any
import triton
from triton.runtime.autotuner import Autotuner
# Entries are (module, outer decorated kernel, exact Triton cache key, config).
# They were selected on Radeon 8060S / gfx1151 for Qwen3.5-35B-A3B training:
# B=4, T=2048, H=HV=32, K=V=128, BF16 q/k/v/beta, FP32 gate,
# no recurrent state, no varlen metadata, and fused Q/K L2 normalization.
_KNOWN_CONFIGS: tuple[tuple[str, str, tuple[Any, ...], dict[str, Any]], ...] = (
(
"fla.modules.fused_norm_gate",
"layer_norm_gated_fwd_kernel",
(
128,
4,
True,
False,
False,
True,
"torch.bfloat16",
"torch.bfloat16",
"torch.bfloat16",
"torch.bfloat16",
"torch.float32",
),
{"kwargs": {"BT": 16}, "num_warps": 16, "num_stages": 3},
),
(
"fla.modules.fused_norm_gate",
"layer_norm_gated_bwd_kernel",
(
128,
4,
True,
False,
True,
"torch.bfloat16",
"torch.bfloat16",
"torch.bfloat16",
"torch.bfloat16",
"torch.bfloat16",
"torch.bfloat16",
"torch.float32",
"torch.float32",
),
{"kwargs": {"BT": 32}, "num_warps": 4, "num_stages": 3},
),
(
"fla.modules.l2norm",
"l2norm_fwd_kernel",
(128, 4, "torch.bfloat16", "torch.bfloat16", "torch.float32"),
{"kwargs": {"BT": 16}, "num_warps": 16, "num_stages": 3},
),
(
"fla.modules.l2norm",
"l2norm_bwd_kernel",
(
128,
4,
"torch.bfloat16",
"torch.float32",
"torch.bfloat16",
"torch.bfloat16",
),
{"kwargs": {"BT": 8}, "num_warps": 8, "num_stages": 3},
),
(
"fla.ops.common.chunk_delta_h",
"chunk_gated_delta_rule_fwd_kernel_h_blockdim64",
(
32,
32,
128,
128,
64,
False,
"torch.bfloat16",
"torch.bfloat16",
"torch.bfloat16",
"torch.bfloat16",
"torch.float32",
"torch.bfloat16",
),
{"kwargs": {"BV": 32}, "num_warps": 4, "num_stages": 1},
),
(
"fla.ops.common.chunk_delta_h",
"chunk_gated_delta_rule_bwd_kernel_dhu_blockdim64",
(
32,
32,
128,
128,
64,
True,
False,
"torch.bfloat16",
"torch.bfloat16",
"torch.bfloat16",
"torch.float32",
"torch.bfloat16",
"torch.bfloat16",
"torch.bfloat16",
"torch.bfloat16",
),
{"kwargs": {"BV": 32}, "num_warps": 4, "num_stages": 1},
),
(
"fla.ops.common.chunk_o",
"chunk_fwd_kernel_o",
(
32,
32,
128,
128,
64,
False,
"torch.bfloat16",
"torch.bfloat16",
"torch.bfloat16",
"torch.bfloat16",
"torch.float32",
"torch.bfloat16",
),
{"kwargs": {"BK": 32, "BV": 32}, "num_warps": 2, "num_stages": 3},
),
(
"fla.ops.common.chunk_o",
"chunk_bwd_kernel_dqkwg",
(
32,
32,
128,
128,
64,
32,
32,
True,
False,
True,
False,
"torch.bfloat16",
"torch.bfloat16",
"torch.bfloat16",
"torch.float32",
"torch.bfloat16",
"torch.bfloat16",
"torch.bfloat16",
"torch.bfloat16",
"torch.bfloat16",
"torch.bfloat16",
"torch.bfloat16",
"torch.float32",
),
{"kwargs": {}, "num_warps": 8, "num_stages": 2},
),
(
"fla.ops.common.chunk_o",
"chunk_bwd_kernel_dv_local",
(
32,
32,
128,
128,
64,
32,
32,
True,
"torch.bfloat16",
"torch.bfloat16",
"torch.float32",
"torch.bfloat16",
"torch.bfloat16",
),
{"kwargs": {}, "num_warps": 8, "num_stages": 2},
),
(
"fla.ops.gated_delta_rule.chunk_fwd",
"chunk_gated_delta_rule_fwd_kkt_solve_kernel",
(
32,
32,
128,
16,
"torch.bfloat16",
"torch.float32",
"torch.bfloat16",
"torch.bfloat16",
),
{"kwargs": {"BK": 32}, "num_warps": 4, "num_stages": 3},
),
(
"fla.ops.gated_delta_rule.wy_fast",
"recompute_w_u_fwd_kernel",
(
32,
32,
128,
128,
64,
64,
64,
False,
"torch.bfloat16",
"torch.bfloat16",
"torch.bfloat16",
"torch.bfloat16",
"torch.bfloat16",
"torch.bfloat16",
"torch.float32",
),
{"kwargs": {}, "num_warps": 8, "num_stages": 4},
),
(
"fla.ops.gated_delta_rule.wy_fast",
"prepare_wy_repr_bwd_kernel",
(
32,
32,
128,
128,
64,
32,
32,
False,
"torch.bfloat16",
"torch.bfloat16",
"torch.bfloat16",
"torch.float32",
"torch.bfloat16",
"torch.bfloat16",
"torch.bfloat16",
"torch.bfloat16",
"torch.bfloat16",
"torch.bfloat16",
"torch.float32",
),
{"kwargs": {}, "num_warps": 4, "num_stages": 2},
),
(
"fla.ops.utils.cumsum",
"chunk_local_cumsum_scalar_kernel",
(4, 32, 64, False, False, "torch.float32", "torch.float32"),
{"kwargs": {}, "num_warps": 1, "num_stages": 3},
),
(
"fla.ops.utils.cumsum",
"chunk_local_cumsum_scalar_kernel",
(4, 32, 64, False, True, "torch.float32", "torch.float32"),
{"kwargs": {}, "num_warps": 2, "num_stages": 3},
),
)
def _unwrap_autotuner(value: Any) -> Autotuner:
for _ in range(8):
if isinstance(value, Autotuner):
return value
value = getattr(value, "fn", None)
if value is None:
break
raise TypeError("decorated kernel does not contain a Triton Autotuner")
def _triton_config(values: dict[str, Any]) -> triton.Config:
return triton.Config(
dict(values["kwargs"]),
num_warps=values["num_warps"],
num_stages=values["num_stages"],
num_ctas=1,
)
def configure_qwen35_fla() -> int:
"""Preload exact gfx1151 FLA autotune winners for Qwen3.5 training.
Only exact Triton cache keys are populated. Different batch geometry,
dimensions, dtypes, recurrent-state modes, or variable-length modes retain
FLA's normal autotuning behavior.
"""
configured = 0
autotuners: dict[tuple[str, str], Autotuner] = {}
for module_name, attribute, cache_key, config_values in _KNOWN_CONFIGS:
identity = (module_name, attribute)
autotuner = autotuners.get(identity)
if autotuner is None:
module = importlib.import_module(module_name)
autotuner = _unwrap_autotuner(getattr(module, attribute))
autotuners[identity] = autotuner
autotuner.cache[cache_key] = _triton_config(config_values)
configured += 1
return configured